LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Abstract

In multilingual retrieval-augmented generation pipelines, embedding models can retrieve relevant documents across languages before reranking. However, existing multilingual rerankers do not consistently prioritize documents in the query language when semantically equivalent candidates are available. We release LAMAR, a language-aware multilingual cross-encoder trained to account for both semantic relevance and language coherence. LAMAR combines English-anchored relevance distillation with preference alignment for language coherence, achieving strong results in controlled language-coherence evaluation and remaining competitive on established multilingual reranking benchmarks.

BibTeX

@article{2026-lamar-open-language-aware-multilingual-alignment-reranker,
  title = {LAMAR: An Open Language-Aware Multilingual Alignment Reranker},
  author = {Seongtae Hong and Youngjoon Jang and Jungseob Lee and Seungyoon Lee and Heuiseok Lim},
  journal = {arXiv preprint arXiv:2607.22042},
  year = {2026},
  month = {7},
  abstract = {In multilingual retrieval-augmented generation pipelines, embedding models can retrieve relevant documents across languages before reranking. However, existing multilingual rerankers do not consistently prioritize documents in the query language when semantically equivalent candidates are available. We release LAMAR, a language-aware multilingual cross-encoder trained to account for both semantic relevance and language coherence. LAMAR combines English-anchored relevance distillation with preference alignment for language coherence, achieving strong results in controlled language-coherence evaluation and remaining competitive on established multilingual reranking benchmarks.},
  url = {https://arxiv.org/abs/2607.22042},
  doi = {10.48550/arXiv.2607.22042}
}